1 / 42

I3: Maximizing Packet Capture Performance

I3: Maximizing Packet Capture Performance. Andrew Brown. Agenda. Why do captures drop packets, how can you tell? Software considerations Hardware considerations Potential hardware improvements Test configurations/parameters Performance results. What is a drop?.

penda
Download Presentation

I3: Maximizing Packet Capture Performance

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. I3: Maximizing Packet Capture Performance Andrew Brown

  2. Agenda • Why do captures drop packets, how can you tell? • Software considerations • Hardware considerations • Potential hardware improvements • Test configurations/parameters • Performance results Sharkfest 2014

  3. What is a drop? • Failure to capture a packet that is part of the traffic in which you’re interested • Dropped packets tend to be the most important • Capture filter will not necessarily help Sharkfest 2014

  4. Why do drops occur? • Applications don’t know that their data is being captured • Result: Only one chance to capture a packet • What can go wrong? Let’s look at the life of a packet Sharkfest 2014

  5. Internal packet flow • Path of a packet from NIC to application (Linux) • Switch output queue drops • Interface drops • Kernel drops Sharkfest 2014

  6. Identifying drops • Software reports drops • L4 indicators (TCP ACKed lost segment) • L7 indicators (app-level sequence numbers revealed by dissector) Sharkfest 2014

  7. When is (and isn’t) it necessary to take steps to maximize capture performance? • Not typically necessary when capturing traffic of <= 1G end device • More commonly necessary when capturing uplink traffic from a TAP or SPAN port • Some sort of action is almost always necessary at 10G • Methods described aren’t always necessary • Methods focus on free solutions Sharkfest 2014

  8. Software considerations - Windows • Quit unnecessary programs • Avoid Wireshark for capturing • Saves to TEMP • Additional processing for packet statistics • Uses CPU • Uses memory over time, can lead to out of memory errors Sharkfest 2014

  9. Software considerations – Windows (continued) • Alternative? Dumpcap • Command-line utility • Called by Wireshark/Tshark for capture • Provides greater control • Dumpcapuifor CLIphobic • “At the limits” example • Dumpcap captured 100% of packets sent • Wireshark captured 68% of packets sent Sharkfest 2014

  10. Software considerations – Windows (continued) • Windows dumpcap buffer tuning • Large buffers are generally good, but… • Increased bandwidth has a tipping point • Write to disk slows significantly • Larger buffers make it worse • Made buffer selection for testing difficult • Best option seemed to be 50MB Sharkfest 2014

  11. Software considerations – Windows (continued) • Dumpcap “slow count” example • Sending 844,600 packets @ .4Gb • Packets take 1.48 seconds to send • 20MB buffer takes ~2.5 seconds to write • 512MB buffer takes ~46 seconds to write • Neither setting captured all packets • Not cosmetic (break out and file is truncated) • Issue disappears at lower bandwidth Sharkfest 2014

  12. Software considerations – Windows (continued) • Video of normal count Sharkfest 2014

  13. Software considerations – Windows (continued) • Video of “slow count” Sharkfest 2014

  14. Software considerations – Windows (continued) • Disable protocols on interface (TAP/SPAN) • Pure TAP/SPAN capture • Only for TAP/SPAN • Prevents OS from attempting to interpret packets • Tested performance with destination MAC set to broadcast address • Result: Captured 100% with protocols disabled,only 40% when enabled • Eliminate performance impact immediately after link up Uncheck everything below Sharkfest 2014

  15. Software considerations – Linux • Quit unnecessary programs • Use tcpdump with 512MB buffer • Ensure libpcap >= 1.0.0 (tcpdump -h) • Watch value of -s flag • No option to disable protocols like Windows • Static (or no) IP for dedicated capture interface • Use XFS with RAID and coordinate stripe sizes Sharkfest 2014

  16. Software considerations – Linux (continued) • Access to development resources? Look at PF_RING • Module/NIC driver combination • Improves capture performance • Included tcpdump wasn’t better than stock • We use the API and it works • Different performance tiers some are free Sharkfest 2014

  17. Software considerations – Linux (continued) • PF_RING • Kernel module/NIC driver combination • Improves capture performance via various methods • Included tcpdump wasn’t better than stock • We use the API and it works • Different performance tiers some are free Sharkfest 2014

  18. Hardware considerations - Storage • 1Gb line rate traffic generates 123-133MB in one second • WD Black 7.2K RPM: 171MB/s • WD Raptor 10K RPM: 200MB/s • If 10Gb is 10X 1Gb… (do the math) • SSD: ~500MB/s • RAM disk is another option Sharkfest 2014

  19. Hardware considerations - CPU • Three considerations • Number of cores • Clock speed • Performance per clock • Clock speed * PPC = Per-core performance • Multicore is good ... • … but per-core performance is better than many cores Sharkfest 2014

  20. Hardware considerations - NIC • Intel (regular NIC) • Drivers more actively maintained • Best PF_RING support • 10G NIC doesn’t help with 1G capture (1G and 10G NICs had the same max bandwidth at or below 1G) • Avoid USB NICs • USB 2.0 is too slow (480Mb/s) • USB 3.0 didn’t perform well Sharkfest 2014

  21. Benchmark methodology • Tested limits of capture configurations at 1G and 10G • For each configuration, increase bandwidth until it fails • Failure is defined as not capturing all packets • Highest performing solutions formed basis for recommendations Sharkfest 2014

  22. Obvious question: Traffic profile? • If not testing for a specific use case, what is the appropriate traffic with which to test? • What mix of TCP/UDP? • What duration, frequency, severity of bursts? • What mix of small/large packets? Sharkfest 2014

  23. (My) Answer: Many copies of a single packet with tests at various packet sizes • Takes Receive Side Scaling out of the picture • Removes buffering from the equation • Tends to be pessimistic Sharkfest 2014

  24. Test configuration • Unicast UDP packet used for (almost) all tests • Packet sizes of 64, 128, 256, 512, 1024, 1500 bytes • Additional CPU overhead for every packet • One second at 1Gb is ~82K 1500 byte packets • One second at 1Gb is ~1.49M 64 byte packets • Number of packets tailored to generate a ~1.5GB capture file • Careful to eliminate disk as a bottleneck Sharkfest 2014

  25. Improving performanceThe ideal • Ideal capture laptop • Fast CPU • Fast storage (SSD RAID) • Dedicated Intel NIC • 10G capability • Perfect except for one issue …it doesn’t exist Sharkfest 2014

  26. Improving performance Thunderbolt • PCIe via a cable (developed by Intel) • Allows use of desktop cards on a laptop • Expensive • Not very widespread (mostly Apple computers) • Other laptop limitations are still a problem Sharkfest 2014

  27. Improving performance Laptop alternative • What level of performance is possible from (relatively) portable commodity hardware? • Packet toaster • Used for all capture testing • Intel i5 4570 desktop CPU (3.6GHz quad-core) • Up to 16GB RAM for RAM disk • Up to 4 SSD in RAID 0 • Cost ~$800 with 8GB RAM, 2 SSDs • Concept: Run without monitor, manage via laptop Sharkfest 2014

  28. Packet Toaster port layout • Intel 1G NIC • Additional 1G NIC for management (SSH/RDP) • 802.11n for capture (Linux) or management • PCIeslot for 10G Sharkfest 2014

  29. Solarflare • Low-latency NIC with stack bypass • Why include it? • Price competitive with other commodity 10G NICs • Works as a regular NIC under Linux, Windows, Mac etc. • Works at 1G also • SolarCapture app for high-performance Linux capture • Hardware/software capture solution • Tested with Packet Toaster and MacBook Pro (via Thunderbolt) Sharkfest 2014

  30. The difference a week makes • At the time of testing, SolarCapture was a free download • Less than a week ago, Solarflare changed licensing tiers; free SolarCapture is no longer available • Pricing is reasonable (in my opinion) but… • …reasonable is relative • …this breaks my original concept of free software • Debated removing results but couldn’t (impacted other results and no time to re-test) Sharkfest 2014

  31. Performance Results Configurations • Wireshark under Windows 7 (SSD) • Dumpcap under Windows 7 (SSD) • Dumpcap under Linux (SSD) • TCPDump under Linux (SSD) • SolarCapture under Linux on MacBook Provia Thunderbolt (RAM) • SolarCapture under Linux (SSD) • SolarCapture under Linux (RAM) Sharkfest 2014

  32. Performance ResultsWireshark vs. Dumpcap (Win 7) Sharkfest 2014

  33. Performance ResultsDumpcap (Win7) - Dumpcap (Linux) – TCPDump (Linux) Sharkfest 2014

  34. Performance ResultsDumpcap (Win7) - Dumpcap (Linux) – TCPDump (Linux) Sharkfest 2014

  35. Performance ResultsTCPDump(Linux) – SolarCapture (SSD) – SolarCapture (RAM) Sharkfest 2014

  36. Performance ResultsDumpcap (Win7) - Dumpcap (Linux) – TCPDump (Linux) Sharkfest 2014

  37. Performance ResultsBy Packet Size Sharkfest 2014

  38. Performance ResultsBy Configuration Sharkfest 2014

  39. Acknowledgements • BATS Global Markets • Guy Harris • Core developer: libpcap, tcpdump and Wireshark Sharkfest 2014

  40. Appendix - Links • Links • http://www.intel.com (Intel NICs) • http://www.ntop.org (PF_RING) • http://www.solarflare.com (SolarCapture) • http://www.tcpdump.org (TCPdump/Libpcap) • http://www.wireshark.org (Wireshark/Dumpcap) • http://www.macsales.com (Thunderbolt enclosure) Sharkfest 2014

  41. Appendix – Packet Toaster Specs • CPU: Intel i5 4570 (3.6GHz quad-core) • Motherboard: Gigabye Z87N-WIFI • RAM: 8GB DDR3 • Storage • Samsung 840 Evo (Operating System) • 2 x Sandisk Extreme in RAID 0 (Capture destination) Sharkfest 2014

  42. Questions Sharkfest 2014

More Related